NER-oriented Chinese clinical text data enhancement method and device

Abstract

The invention discloses an NER-oriented Chinese clinical text data enhancement method and device. Unlabeled data and labeled data subjected to label linearization processing are obtained through data preprocessing. Unlabeled data is used, part of information in a text is masked, the masked part is predicted on the basis of reserved information, meanwhile, an entity word level judgment task is introduced, and fragment-based language model pre-training is carried out; a plurality of decoding mechanisms are introduced in a fine tuning stage, a relationship between a text vector and text data is obtained based on a pre-trained fragment-based language model, linearized data with entity labels are converted into the text vector, text generation is carried out through forward decoding and reverse decoding in a prediction stage of a text generation model, and the text generation model is used for generating text data. The method comprises the following steps of: analyzing a label to obtain enhanced data with annotation information; according to the method, the data diversity is further improved, and the quality of the enhanced data is also improved, so that the model can generate more high-quality enhanced data.

Description

technical field [0001] The invention relates to the field of text data enhancement, in particular to an NER-oriented Chinese clinical text data enhancement method and device. Background technique [0002] Named entity recognition task is a basic task in the field of natural language processing. It is a kind of sequence labeling problem. Similar to classification tasks, each unit in a text sequence (Chinese named entity recognition task is usually based on single words or subwords). processing) for category judgment, and the judgment results usually include various categories such as "non-entity", "entity beginning word", "entity middle word", "entity ending word", among which, the entity-related type will be determined according to the entity to be predicted. Types vary. [0003] With the advancement of medical informatization construction, the amount of medical text data shows an explosive growth trend. The extraction and utilization of information in unstructured medical ...

AI Technical Summary

Problems solved by technology

However, the method of enhancing through language model generation is mostly based on single characters or subword sequences for text prediction, and most medical entities are composed of fixed semantic units. When the general method is directly applied to the medical field, it will be ignored. Losing the unique semantic characteristics of medical entities may result in generated data that may not conform to the characteristics of medical terms or violate medical logic, thereby affecting the accuracy of the NER model

[0009] Common generative models mostly use left-to-right decoding, which can only use the historical information that has been generated, but cannot use the future information that has not yet been generated, resulting in generated There is a certain degree of paranoia in the sample; at the same time, as the generation sequence becomes longer, the single-direction generation method is likely to aggravate the problem of error accumulation: for example, if an unreasonable vocabulary is generated somewhere in the middle, it will bias the follow-up prediction results and affect the quality of the overall generated sample

Images